Artificial intelligence in sepsis early prediction and diagnosis using unstructured data in healthcare

Kim Huat Goh,Le Wang,Adrian Yong Kwang Yeow,Hermione Poh,Ke Li,Joannas Jie Lin Yeow,Gamaliel Yu Heng Tan

DOI: https://doi.org/10.1038/s41467-021-20910-4

IF: 16.6

2021-01-29

Nature Communications

Abstract:Abstract Sepsis is a leading cause of death in hospitals. Early prediction and diagnosis of sepsis, which is critical in reducing mortality, is challenging as many of its signs and symptoms are similar to other less critical conditions. We develop an artificial intelligence algorithm, SERA algorithm, which uses both structured data and unstructured clinical notes to predict and diagnose sepsis. We test this algorithm with independent, clinical notes and achieve high predictive accuracy 12 hours before the onset of sepsis (AUC 0.94, sensitivity 0.87 and specificity 0.87). We compare the SERA algorithm against physician predictions and show the algorithm’s potential to increase the early detection of sepsis by up to 32% and reduce false positives by up to 17%. Mining unstructured clinical notes is shown to improve the algorithm’s accuracy compared to using only clinical measures for early warning 12 to 48 hours before the onset of sepsis.

multidisciplinary sciences

What problem does this paper attempt to address?

This paper aims to solve the problems of early prediction and diagnosis of sepsis. Sepsis is one of the leading causes of death in hospitals, and its early prediction and diagnosis are crucial for reducing mortality. However, because many symptoms of sepsis are similar to those of other less - severe diseases, early identification of sepsis is challenging.

To meet this challenge, researchers have developed an artificial intelligence model named SERA (Sepsis Early Risk Assessment) algorithm. This algorithm uses structured data (such as vital signs, test results and treatment information in electronic medical records) and unstructured clinical notes (such as doctor records in free - text form) to predict and diagnose sepsis.

Specifically, the research objectives of the paper include:
1. **Improve prediction accuracy**: By combining structured data and unstructured clinical notes, improve the early prediction accuracy of sepsis.
2. **Extend the warning time**: Achieve a 12 - to 48 - hour warning for sepsis in advance, providing more preparation time for clinicians.
3. **Reduce the false positive rate**: Compared with the doctor's prediction, reduce the false positive rate and improve the practicality of the algorithm.

The research results show that the SERA algorithm can accurately predict the occurrence of sepsis 12 hours in advance on an independent test set, with an AUC value reaching 0.94, and the sensitivity and specificity are 0.87 respectively. In addition, compared with the doctor's prediction, the SERA algorithm can increase the number of early - detected sepsis cases by 32% and reduce the false positive rate by 17% at the same time. These results indicate that the SERA algorithm has significant advantages in improving the early prediction and diagnosis of sepsis.